Research map / Detection and segmentation / CNN features
DETRs with Collaborative Hybrid Assignments Training
Zhuofan Zong, Guanglu Song, Yu Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add collaborative one-to-many auxiliary head and customized positive queries | Deformable-DETR · COCO val 12 epochs | AP: 37.1 -> 42.9 | Improved |
| add Co-DETR to DINO-Deformable-DETR Swin-L | DINO-Deformable-DETR · COCO val 36 epochs | AP: 58.5 -> 59.5 | Improved |
| add auxiliary head with one-to-many assignment (ablation) | Deformable-DETR · COCO val 36 epochs | AP: 43.3 -> 46.2 | Improved |
| add customized positive queries generation (ablation) | Deformable-DETR · COCO val 36 epochs | AP: 46.2 -> 46.8 | Improved |
| increase number of different auxiliary heads from 3 to 6 (ablation) | Deformable-DETR++ · COCO val 12 epochs | AP: 49.5 -> 48.9 | Hurt |
| increase auxiliary heads from K=1 to K=2 (ATSS + Faster-RCNN) (ablation) | Deformable-DETR++ · COCO val 12 epochs | AP: 48.7 -> 49.5 | Improved |
| use K=2 versus K=1 with Deformable-DETR 36 epochs (ablation) | Deformable-DETR · COCO val 36 epochs | AP: 46.8 -> 46.5 | No clear effect |
| choose any one-to-many head (RetinaNet to ATSS) as auxiliary (ablation) | Deformable-DETR · COCO val 36 epochs | AP: 43.3 -> 46.1 to 46.8 | Improved |
| train baseline Deformable-DETR longer instead of using Co-DETR (ablation) | Deformable-DETR · COCO val | AP: 46.0 (100 ep) -> 45.9 (150 ep) vs 46.8 with Co-DETR at 36 ep | No clear effect |
| use two ATSS heads versus one (ablation) | Deformable-DETR++ · COCO val 12 epochs | AP: 48.7 -> 49.2 | Improved |
| scale Co-DETR to ViT-L backbone with Objects365 pretraining | Co-DINO-Deformable-DETR · COCO test-dev | AP box: 65.5 -> 66.0 (vs InternImage-G) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Deformable-DETR — Deformable DETR: Deformable Transformers for End-to-End Object Detection (ICLR 2021) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Teach-DETR: Better Training DETR With Teachers (TPAMI 2023 · similarity 0.87)
- DAC-DETR: Divide the Attention Layers and Conquer (NeurIPS 2023 · similarity 0.87)
- MS-DETR: Efficient DETR Training with Mixed Supervision (CVPR 2024 · similarity 0.86)
- Salience DETR: Enhancing Detection Transformer with Hierarchical Salience Filtering Refinement (CVPR 2024 · similarity 0.85)
- Order Matters: Competition-Guided Query Ordering for RNN-Based Object Detection (NeurIPS 2026 · similarity 0.84)
- Hybrid Proposal Refiner: Revisiting DETR Series from the Faster R-CNN Perspective (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
